Data-Driven Design and Control Techniques for Soft Hands with Embedded Rigid Constraints

Soft manipulation goes beyond traditional manipulation paradigms, introducing innovative strategies to facilitate the robust interaction between robotic hands and their sorroundings. The approach known as Environmental Constraint Exploitation (ECE), harnesses the features of the environment to enhance grasp robustness while minimizing planning efforts. The pivotal factor enabling ECE is the inherent compliance of soft hands, which allows them to comply and adapt to environmental features. The foundational step toward achieving stable grasps involves emulating how humans naturally exploit environmental constraints. However, the accurate detection of environmental constraints can be complex or not possible in many cases. A possible solution to deal with this issue is to embed, directly in the hand structure, purposefully designed parts that work themselves as ”embedded constraints”. This paradigm shift, for example, can be obtained by endowing a robotic gripper with an additional soft-rigid palm that can slide over and in between flat surfaces. This Thesis presents design and control techniques that allow the development and use of soft-rigid grippers with embedded constraints to effectively exploit environmental constraints and robustly grasp a variety of objects. Three distinct approaches are employed for the control: the first calculates the pre-grasp pose of the gripper through analytical optimization, the second relies on data-driven approaches based on human demonstrations, while the third is based on Deep Reinforcement Learning. Regarding the design, this Thesis shows that the application of optimization techniques combined with machine learning algorithms can support the automated design of novel gripperss, which can work efficiently. The proposed control and design strategies aim at broadening the application of robotic manipulation beyond industrial settings, enabling robotic systems to operate effectively in unstructured and previously unseen scenarios.